Models Overview

At Kid Smart AI, we believe

1. Transparency is the key to building trust with both our partners and parents and guardians
2. Child developmental experts must be involved from the beginning
3. Continuous improvement is needed, so feedback from partners is invaluable

We use a both pretrained and proprietary models to ensure best possible outcome for each application.

We currently leverage the following open source pre-trained models:

Purpose	Model Name	Pretrained Model Developer	Model Host	Model Card
Speech Recognition	Whisper V3	OpenAI	Kid Smart AI	model card
Text Generation	llama-3	Meta	Kid Smart AI	model card
Image Generation	Stable Diffusion XL (SDXL)	Stable Diffusion	Kid Smart AI	paper

We leverage pre-trained models rather than building our own models because, for instance, teaching a model to read and write english is very expensive (it can cost millions of dollars to train these models). Due to the cost of training, some companies then "open-source" their models so that individuals and businesses can modify them for their own needs.

For each of these pretrained models, we modify them heavily for the young child use case. This includes all of the following:

• Use case specific modification of the model inputs or representations (eg tokenizer modification or ReFT)
• Adjustment of the model weights (techniques like RLHF, continued pre-training or LoRA)
• Add additional layers and/or remove layers

When no pre-trained alternatives exist, we train our own proprietary models, like below.

Purpose	Model Name	Model Developer	Model Host	Model Card
Fluency Error Classification	fluency_v1	Kid Smart AI	Kid Smart AI	In progress
Proununciation	pronunciation_v1	Kid Smart AI	Kid Smart AI	In progress
Word Recognition	recognition_v1	Kid Smart AI	Kid Smart AI	In progress

We use other models, like Open AI GPT-4, Anthropic, Meta Wav2Vec, and OpenAI Whisper V2 for internal purposes.